Design page for Promises / Futures with callbacks.
Motivation
Clojure's future
and promise
provide implementations of deref
that block the calling thread. There is currently no facility to invoke callback functions on values when they become available. Asynchronous styles of programming are not possible without this facility, especially in single-threaded runtimes such as ClojureScript.
Proposed APIs
Low-Level Primitives
promise
and future
will return objects implementating a new protocol, INotify
, supporting a function attend
with two arities:
attend [promise fn] [promise fn executor]
This function attaches a callback to the promise/future which will be invoked on the executor when the promise is delivered with a value. (attend, transitive verb, "occur with or as a result of")
Promises also implement the new protocol IDeliver
, supporting two functions:
deliver [promise value] fail [promise exception]
The deliver
function provides a value to the promise, invokes any pending callbacks, and releases any pending derefs. The fail
function delivers an exception to the promise (see "Exceptions" below).
The behavior of deref
on futures/promises does not change, except that a failed promise may throw an exception (see below).
Higher-Level Functions & Macros
Two macros facilitate composing promises:
then [promise binding-form & body]
recover [promise symbol & body]
The then
macro waits for the promise to be delivered successfully and executes body
in a callback with the value of the promise bound to binding-form
. It returns a new promise which will be delivered with the value of body. If the promise fails or if body
throws an exception, then
propagates the exception to the returned promise.
The recover
macro catches exceptions thrown from a failed promise, executes body
with the exception bound to symbol, and returns a new promise which will be delivered with the value of body. If the promise does not fail, recover
propagates its value to the returned promise. This behavior is intended to be similar to try/catch
.
These macros can be composed with the threading macro:
The then
and recover
macros are implemented in terms of two functions, then-call
and recover-call
, which take functions as arguments.
Execution Context
If not supplied, executor
defaults to a dynamic Var, *callback-executor*
, resolved when the callback is created, that is, at the time attend
was called. This Var is initially set to the Agent send-off thread pool executor, and can be globally reset or thread-bound by applications.
Running Callbacks Locally
If callbacks are known to be fast, it may be more efficient to run them directly on the thread that invoked deliver
. This is enabled by calling attend
with a nil
executor (or binding *callback-executor*
to nil
).
Guava provides a similar capability with sameThreadExecutor. It is somewhat related to the JDK's CallerRunsPolicy for rejected execution.
Further Analysis
Delivery
The APIs in this proposal can be delivered either as a library or as an extension to the Clojure language.
Extending the Clojure language has the benefits of enhancing the existing promise
and future
in a standard way. Having a standard mechanism for callbacks at the language level will facilitate sharing callbacks across different libraries and frameworks.
Exceptions
A callback function (really, any function) can do one of three things:
- Return a value
- Throw an exception
- Never return at all
Failing to return is almost certainly a bug anywhere execpt on the main application thread. I assume that returning a value indicates a successful execution, while throwing an exception indicates failure. However, it is also possible to return an exception object as a value, which does not necessarily indicate failure. (For example, a logging framework might handle exceptions as values.)
Given that throwing an exception is different from returning a value, promises need to handle it on a different path. This is the fail
function, which acts like deliver
but takes an exception as its argument and puts the promise in a "failed" state.
When it comes time to invoke callback functions, there are several possible ways to handle failure:
- Provide separate callback attachment points for success and failure
- Pass the exception to the callback as a normal value
- Wrap the value or exception in a union type
Method 1 leads to a proliferation of error-handling functions. Method 2 makes it impossible to tell if the exception object was thrown or delivered as a normal value. Method 3 is more natural in a statically-typed language (Scala does this). In both 2 and 3 the callback function must always check whether its argument is a normal value or an exception.
I have chosen a variant of 3 by using the promise itself as the argument to the callback function. The callback is responsible for calling deref
on the promise. If the promise is in a failed state, dereferencing it will throw the exception. Callback functions can use standard try/catch blocks to handle the exception or allow it to propagate. Macros such as on
handle propagation automatically and allow users to program in terms of values.
Delivering a promise A as a value to another promise B causes B to attend
on A. Whenever A is delivered/failed, B is delivered/failed with the same value. This makes it convenient to implement multiple try/retry operations with promises, as in this example:
Java Interop
Promises implement java.util.concurrent.Future. Extending them to similar APIs such as Guava's ListenableFuture is trivial.
Cancellation
There is no API to remove or cancel a callback created with attend
. However, it is possible to deliver a value to the promise returned by the then
macro before the underlying promise is delivered. In that case, the callback function will not be executed. There is an implicit race condition with cancellation, so this cannot guarantee that the callback function will never be executed at all.
Futures can still be cancelled with future-cancel
, which attempts to stop the Future's thread.
Calling future-cancel
on a promise is equivalent to calling fail
with an instance of java.util.concurrent.CancellationException.
Progress Reporting
Promises do not provide an API to report partial progress during a long-running operation. Progress reporting can be handled through other means such as watches.
ClojureScript
Promises can be implemented in ClojureScript, subject to the following limitations:
- There is no notion of executors apart from "execute immediately" or "execute asynchronously" (e.g., with
setTimeout
) deref
on an unrealized promise cannot block; instead it throws an exception
Discussions
- Promise/Futures with callbacks (Clojure-dev group)
- lamina and channels-driven concurrency with clojure (Clojure group)
- reactive programing (dev.clojure.org)
- can a better Future be exposed by API? (Datomic group)
Implementations
- cljque promises (Stuart Sierra)
References / Related Work
- Java
- Listenable Futures in Google Guava
- ListenableFuture in Async HTTP Client
- Scala
- JavaScript
- CommonJS Promises/A
- Promises/A%2B (extension of Promises/A)
- You're Missing the Point of Promises
- Q
- Angular JS promise/deferred
- C# / .NET
- Python
Comparison to Guava ListenableFuture
Guava's ListenableFuture takes the same basic approach of passing a Runnable and an Executor for the callback. It should be trivial to extend the INotify and IDeliver protocols to Guava's ListenableFuture and SettableFuture.
The implementation in Guava is more sophisticated than the one in cljque, using an AbstractQueuedSynchronizer instead of monitors.